What Happens When Institutions Can Finally See Themselves

When an institution cannot see itself — cannot measure its own operations, track its own outputs, account for its own resources — it cannot improve. It can only react. Problems surface as crises. Decisions get made on instinct. Resources disappear into opacity. The institution drifts.

This is the condition of most of Africa's public institutions today. Not because the people inside them are incompetent. But because they have been operating without the instruments that would let them see. I have called this the missing layer elsewhere: the absence of an operational data infrastructure that would let an institution observe its own functioning in real time. This paper is about what changes, concretely, when that layer gets built — and about how much of that change this paper can actually claim to have measured, as opposed to observed.

1. What this paper can show, and what it can only argue

This paper makes a narrower claim than its title might suggest, and it is worth being precise about the claim before making it.

What it argues from. Direct fieldwork across water utilities in Kenya, conducted while building AccessWASH, including the operations-manager quotation in §2. Deployment metrics from the first wave of AccessWASH utilities, 2021–2023 — Flux's own operational data, not independently audited. Kenya's water-sector regulator's own published performance data for the sector as a whole.¹ One randomized field experiment on whether making service-provider performance visible to the people who depend on it changes provider behaviour and client outcomes, from a different country and a different mechanism, used here as the closest controlled evidence this paper has for its causal claim rather than as a direct measurement of it.² A peer-reviewed framework for why information systems built for one institutional context fail when transplanted into another.³

What it does not argue from. No independent audit of the AccessWASH before/after table in §3 exists. No randomized or quasi-experimental study isolates the specific causal effect of internal operational visibility (as opposed to external community monitoring, which is the mechanism in the one controlled study this paper cites) on an African public institution's performance. The claim that visibility redistributes power inside an institution is argued from observation, not measured with an instrument that could falsify it.

What it does not claim. It does not claim visibility is sufficient — §8 and the companion paper on the operating workforce are explicit that a dashboard which reveals eleven overdue reconnections does not reconnect anyone. It does not claim the AccessWASH numbers in §3 generalise to every utility, sector, or country; they are one vendor's own account of one wave of its own deployments. It does not claim visibility is costless to the people being made visible; §8 takes the "this feels like surveillance" objection seriously rather than dismissing it.

What would refute it. Given in full in §10. The sharpest version: if institutions that gain comparable operational visibility show no measurable improvement in the specific metrics in §3 — revenue collection, fault resolution time, reporting latency — once that comparison is made independently of Flux's own data, the mechanism this paper describes is not doing the work this paper says it is doing.

2. What "seeing" actually means

There is a tendency, when discussing institutional technology, to reach immediately for the glamorous — dashboards, AI, predictive analytics. But the problem Africa's institutions face is more fundamental than that. Most of them do not have a reliable answer to basic questions:

• How many customers does this water utility serve?
• Of those, how many paid last month?
• Of those who didn't pay, how many have been disconnected?
• How many reconnections are pending?
• What is the average time between a reported fault and its resolution?

These are not sophisticated questions. They are the operational basics — the floor of institutional visibility. And in most African utilities, constituency offices, and community health structures, they are unanswerable without significant manual effort, if they are answerable at all.

This is the gap. Not AI. The gap is structured data about what the institution actually does, updated in real time, accessible to the people who need to act on it.

"I cannot tell you how many connections we have right now. I would have to go through the ledgers manually. That would take days. The number changes daily." <span slot="attribution">— Operations Manager, mid-sized Kenyan water utility, 2023</span>

When we first started working with water utilities in Kenya as part of the research that became AccessWASH, we expected to find outdated technology. We found something more surprising: no technology at all, in the sense that mattered. Billing was done in spreadsheets emailed between departments. Fault reports came in on WhatsApp. Reconnection orders were handwritten and physically delivered to field teams. Asset records — the map of the network, the location of each valve and meter — existed in one engineer's mental model, occasionally committed to paper.

This is not a technology problem in the sense that "they need better software." It is an instrumentation problem. The institution has not been equipped with the instruments that let it observe itself.

Kenya's own water-sector regulator makes a version of this point in its most recent sector-wide performance report, in the specific language of non-revenue water (NRW) — the share of water a utility produces that is never billed. Nationally, NRW rose from 43% to 44% between the two most recent reporting years, against a regulatory benchmark of 25%; the gap between the two, once allowable losses are stripped out, comes to roughly 203 million cubic metres a year, valued at KES 11.9 billion.¹ The same report states plainly that closing part of that gap is a data problem as much as an infrastructure one: "the use of technology is essential to boosting data-driven decision-making, monitoring compliance, and improving the delivery of water services."¹ That is the regulator's own diagnosis, not Flux's marketing copy.

It is worth being precise about what that 44% figure is and is not. It is a national average across utilities of very different sizes and conditions, reported by the regulator rather than independently audited by this paper. The same report gives a county-level range for the same period — from 28% (Nyeri county, the best-performing) to 75% (Marsabit county, the worst) — which is the clearest evidence that this is not a uniform national condition but a wide distribution, and that "Africa's water utilities lose roughly 40%" flattens a range that spans nearly 50 percentage points.¹

3. What changes when you can see

When an institution gains operational visibility — when it can finally see itself — three things happen in sequence, on the account this paper is making. The first of the three is the one closest to being something this paper can show rather than argue.

First, problems become legible. Before visibility, a water utility that loses a large share of treated water to non-revenue losses simply experiences that loss as a vague pressure: revenue doesn't add up, the numbers don't balance, there is a feeling of scarcity. After visibility, non-revenue water becomes a number, measured on a cadence shorter than a month. It has a location. It has a pattern. It becomes something that can be acted on. This is the least contestable of the three claims, because it is close to a definitional truth: a number you did not have before is, trivially, now something you have. The interesting question — taken up in §4 — is whether having the number changes what the institution does next, which is a behavioural claim, not an informational one.

Second, accountability becomes possible. This is the uncomfortable change, and the one that generates the most institutional resistance. When performance is measured, performance can be evaluated. The constituency office that processes applications in fourteen days instead of the mandated three is now visible. The field team that closes fault reports without actually visiting the site is now visible. Visibility redistributes power. It shifts information from being held by individuals to being held by the institution. The closest controlled evidence this paper has for this specific mechanism — that visibility changes provider behaviour and downstream outcomes — comes from a different country, a different sector, and a different direction of visibility: a randomized field experiment in rural Uganda found that giving communities structured information about their local health clinic's performance, and a forum to act on it, produced a 33% reduction in under-five mortality and a 20% increase in general outpatient utilization within a year.² That is community-facing accountability, not the internal operational visibility this paper is mostly about, and §8 treats the difference as a real one rather than papering over it — but it is the best controlled demonstration this paper can point to that visibility changing what a service provider does is not merely a plausible story.

Third, learning becomes structural. An institution that can see itself can compare this month to last month. It can run an experiment — change a process, measure the effect. It can train new staff against documented procedures rather than tribal knowledge. Institutional knowledge stops being locked inside individual heads where it disappears when staff leave. It becomes embedded in systems where it compounds.

Selected operational metrics: before and after digital instrumentation at a Kenyan water utility, 2021–2023. Flux's own deployment data, not independently audited.

These numbers are drawn from the first wave of AccessWASH deployments at Kenyan water utilities between 2021 and 2023. They are Flux's own account of its own product's effect, measured by Flux, reported by Flux. No external auditor has verified this table, and no comparison utility that did not adopt AccessWASH was tracked over the same period to rule out other explanations — a general improvement in management attention, a change in leadership, a tariff adjustment — for the same gains. The table is evidence that something changed at this utility during this window, in the direction the thesis predicts. It is not independent proof that visibility caused it.

4. Why visibility alone is not the whole mechanism

It would overstate this paper's own argument to stop at §3. A dashboard is informative, not actuating — a theme the companion paper on the operating workforce develops at length, and this paper borrows its sharpest line rather than re-deriving it: "the missing layer, fully built, produces an institution that can finally watch itself fail in high resolution." A constituency office that can now see eleven overdue bursary applications has gained information. Whether those eleven get processed depends on whether anyone has the hours, the authority, and the incentive to act on what the dashboard shows.

This matters for how §3's three claims should be read. The first — problems become legible — is close to automatic once the instrumentation exists. The second and third — accountability becomes possible, learning becomes structural — are conditional. They happen when an institution's management chooses to use the new visibility to change incentives and processes. They do not happen automatically, and this paper has not measured how often, across how many institutions, that choice gets made once the dashboard exists. That is one of this paper's own open questions, registered in Appendix B rather than assumed away.

5. Why visibility is so hard to achieve

If the gains are so clear, why don't all institutions invest in operational visibility? The barriers are real.

The data collection problem. Building the instruments requires knowing what to measure. This sounds obvious, but in practice it requires deep sectoral knowledge. Water utilities and constituency offices are not generic organisations — they have specific workflows, specific constraints, specific failure modes. Generic software — imported enterprise systems designed for European municipalities — rarely fits. There is a well-documented, peer-reviewed name for why: a "design-actuality gap" between what an information system assumes about an organisation's information, technology, processes, staffing, management structures and other resources, and what that organisation actually has. The wider the gap along any of those dimensions, the more likely the system fails outright or survives only as a shell nobody uses.³ The instrumentation has to be built for the specific institution, not adapted from one built for a different one.

The connectivity problem. Many of the operations that need to be tracked — field service, meter reading, community health visits — happen in places with unreliable or no internet access. An instrumentation system that requires constant connectivity will fail exactly when it is most needed. Offline-first architecture is not an optional feature. It is a prerequisite — the subject of a companion paper in this series, which this one does not re-argue.

The adoption problem. An institution that has operated on manual processes for decades has a deeply embedded workflow. The forms, the ledgers, the WhatsApp groups — these are not just tools. They are the way things are done. Replacing them requires training, and training requires sustained engagement. A software deployment that ends with a handover and a manual does not change how an institution operates. Institutions change through practice, not documentation. The same design-actuality literature puts a rough number on how often this goes wrong across the wider field of public-sector information systems in developing and transitional economies: one practitioner's own working estimate, built from an informal 2002 poll of field experts and roughly forty case write-ups rather than a probability sample, put total failure at about 35%, partial failure at about 50%, and clean success at about 15%.⁴ That estimate is the author's own description of a "dirty reality," offered honestly as a working figure rather than a measured rate, and this paper repeats it on the same terms rather than upgrading its certainty.

The trust problem. In many African public institutions, data visibility has historically been associated with external audit and accountability mechanisms — donors, regulators, oversight bodies. To the institution's own staff, being made visible feels like being watched. This is not a minor objection, and §8 takes it up directly rather than as a coda here.

6. The translation layer

What we have learned, building AccessWASH and BungeConnect, is that the gap between "the technology exists" and "the institution can see itself" is not a technical gap. The technology does exist. The gap is translation — the same gap Heeks's design-actuality framework names more formally in §5.

Translation means understanding the institution's actual workflow before proposing to change it. It means knowing that the field technician in Nakuru who records faults on paper does so because the mobile app crashes when the network drops — and building offline-first before expecting adoption. It means recognising that the constituency office manager who keeps constituent records in a physical ledger does so because every digital system previously deployed has disappeared when the project funding ended — and building for sustainability before expecting trust.

Translation is expensive. It requires being inside the context, not outside it. It requires sustained presence — not a deployment visit, but an embedded relationship over years. It is the thing that most technology interventions in African institutions skip, because it is hard to invoice and hard to put in a slide deck.

It is also the only thing that makes the technology work. A design-actuality gap narrow enough not to fail, on Heeks's own framework, is not produced by better software alone — it is produced by someone doing the translation work of closing the gap along each of its seven dimensions, for this institution, not a generic one.³

7. What comes next

The first generation of digital instrumentation for African institutions is underway. The utilities that have deployed AccessWASH can now see themselves, on their own unaudited account in §3. The constituency offices piloting BungeConnect can track constituent requests from intake to resolution. These are early deployments — but they demonstrate that the vision is achievable, not that it is proven at scale.

What comes next is harder: making these systems self-sustaining, building the institutional capacity to maintain and evolve them without external technical support, and proving the model across enough deployments — independently measured, not only self-reported — that it becomes the default rather than the exception. The tools themselves are not the scarce resource; the translation work that fits them to a specific institution is.

The technology exists. The translation work is the work. Visibility is the first step — and the step that makes everything else possible, on the condition, named in §4, that someone acts on what it reveals.

— Ken Ruto

8. Objections I take seriously

"Flux has a commercial interest in this conclusion." Flux built AccessWASH, and the deployment table in §3 is Flux's own account of Flux's own product's effect. I am not going to pretend the incentive to tell a flattering story isn't there. The honest mitigation is partial: the WASREB sector-wide figures in §2 and the Björkman and Svensson result in §3 are both independent of Flux and would show the same thing whether or not Flux existed. What is not independent is the specific causal leap from "visibility plausibly improves institutional performance, per a different country's different mechanism" to "AccessWASH's own table in §3 proves it did so here" — that inferential step rests on Flux's own, unaudited measurement.

"Visibility is surveillance, not service." James C. Scott's history of state legibility campaigns argues that making a population legible to a central authority — standardised names, cadastral surveys, population registers — has historically served the authority's power to tax, conscript, and control at least as often as it has served the people being counted.⁵ This is a serious historical argument, and it does not disappear just because this paper's subject is an institution observing itself rather than a state observing its citizens. The honest response is narrower than a rebuttal: Scott's subjects did not choose to be made legible and did not control what the resulting records were used for. A water utility's own field technicians and billing staff, in the cases this paper describes, are supposed to be the primary beneficiaries of the resulting visibility — faster fault resolution, a system that no longer loses what they reported. Whether that distinction holds in practice, and does not quietly slide into a tool used only against frontline staff rather than for them, is not something this paper has measured, and the "trust problem" named in §5 is this paper's own admission that the slide is a live risk, not a solved one.

"The AccessWASH table proves correlation, not the mechanism this paper claims." Correct, and stated plainly in §3. No comparison group, no independent audit, no ruling-out of concurrent changes in management or funding. This is the single largest evidentiary gap in this paper, and §10's sharpest falsification condition is built around closing it rather than around disputing that it exists.

"The Björkman and Svensson result is the wrong kind of evidence for this paper's claim." Also a fair objection. Their intervention changed what communities knew about a clinic they do not control from the inside; this paper is mostly about what an institution knows about itself. The mechanisms — external social pressure versus internal management feedback — are not obviously the same mechanism, and this paper treats their result as the best available analogy for "visibility changes provider behaviour," not as a direct measurement of internal operational visibility's effect.

9. Limitations, and what this paper does not claim

The AccessWASH table in §3 is Flux's own deployment data, reported by Flux, about Flux's own product — not an independent audit.

It does not claim visibility alone produces the second and third changes in §3; §4 is explicit that accountability and structural learning require an institution's management to act on what visibility reveals, and this paper has not measured how often that happens.

It does not claim the Björkman and Svensson finding is a direct measurement of internal institutional self-visibility; it is the closest controlled evidence this paper has for a related but distinct mechanism, and §8 states the distinction rather than blurring it.

It does not claim Kenya's national 44% non-revenue-water average applies uniformly; §2's county-level range (28%–75%) is the more honest picture.

And it does not claim the "surveillance, not service" objection in §8 is resolved; it names a live risk this paper has not independently checked against evidence from its own deployments.

10. What would falsify this

Stated as commitments, in decreasing order of how much damage each would do to this paper's claims.

An independent, non-Flux measurement of utilities that adopt comparable operational visibility tools shows no improvement in revenue collection, fault resolution time, or reporting latency relative to utilities that do not, controlling for other changes over the same period. That would directly contradict this paper's central claim and the right response would be withdrawal of the causal claim in §3, not a softer restatement of it.

A rigorous audit of AccessWASH's own deployment data finds the before/after numbers in §3's table were materially different from what Flux reported. That would be the sharpest indictment of this specific paper's own evidentiary practice, independent of whether the broader thesis about visibility survives elsewhere.

Staff-level evidence from Flux's own deployments shows operational visibility tools being used primarily to discipline or surveil frontline workers rather than to help them do their jobs. That would vindicate the "surveillance, not service" objection in §8 and require this paper's framing of visibility as something that serves the people being made visible to be substantially walked back.

A well-resourced, well-governed African institution with ample staff and funding is shown to remain unable to answer the basic operational questions in §2 despite having invested in modern information systems. That would suggest the missing layer is not primarily an instrumentation problem at all, undermining this paper's and its companion papers' shared diagnosis.

Appendix A — Evidence table

Appendix B — Citation register

CITATION-NEEDED — independent measurement of visibility's effect on African institutional performance. Appendix A row 9. No study independent of Flux was found measuring the specific causal claim in §3 for an African public institution; this paper relies on Flux's own deployment data plus an analogous but distinct controlled result from a different country and mechanism (row 4).

CITATION-NEEDED — rate at which visibility translates into institutional action. Appendix A row 10. §4 names this as an open, unmeasured question rather than assuming the answer.

CITATION-NEEDED — whether Flux's own visibility tools have been used against, rather than for, frontline staff. Appendix A row 11. §8's "surveillance, not service" objection is answered narrowly and provisionally, not with evidence from an audit of actual use.